arch-strix-halo-pkgs

Arch packaging for a Strix Halo local inference stack on Arch and CachyOS.

This repository turns a fast-moving ROCm, vLLM, llama.cpp, and Lemonade setup into something you can inspect, rebuild, publish, install, and repair with normal Arch package tools. It is a source and policy workspace, not a binary package release.

Note

The packages here are tuned and validated on a Strix Halo reference host with a Radeon 8060S (gfx1151). If stock Arch, CachyOS, AUR, or vendor packages already give you the behavior you need, start there. This repo is for the narrower case where you want the Strix Halo-specific package stack and its patch history.

Why This Exists

Local inference stacks are easy to assemble once and hard to keep coherent. This repo keeps the moving parts visible:

• ROCm and TheRock split packages for the gfx1151 runtime and compiler lane
• source-built and rebuilt Python packages for the vLLM runtime closure
• llama.cpp HIP and Vulkan backends packaged as system-managed runtimes
• Lemonade server/app packages wired to those packaged backends
• package metadata, patch files, freshness checks, and validation notes that a maintainer can audit without chat history

The goal is not to hide complexity. The goal is to put each decision where it can be reviewed: source patches beside packages, update policy under docs/, package state in maintainer docs, and host validation in tracked scenarios.

Who This Is For

• Strix Halo users on Arch-family systems who want a local package repo for ROCm, vLLM, llama.cpp, and Lemonade instead of one-off build directories.
• Packagers and maintainers who want to compare the local stack against Arch, CachyOS, AUR, upstream ROCm, and Blackcat Informatics recipe inputs.
• Future agents continuing this repo who need durable package context, current blockers, and exact update rules.

This is not a turnkey installer for every AMD GPU. It is a packaging workspace for one hardware lane, with enough documentation to explain how the lane is assembled and where it diverges.

What It Packages

The stack currently includes these major layers. The candidate column describes the 2026-08 issue-primary convergence line, not installed or adopted package state. Its W0-W2B links point to execution work units nested under the W0, W1, W2A, and W2B wave issues under execution umbrella #95. The W2A runtime work units #109 and #111 remain distinct. W2B work unit #113 intentionally owns both llama.cpp HIP/Vulkan closure and the coherent Lemonade family build because those source and package gates are coupled.

Layer	Package examples	2026-08 candidate line
ROCm foundation	therock-gfx1151, rocm-gfx1151, split ROCm libraries	TheRock 7.14 in W1
Host math and Python	aocl-utils-gfx1151, aocl-libm-gfx1151, python-gfx1151	CPython 3.14.7 in W1; AOCL remains governed by current package evidence
ML runtime	python-pytorch-opt-rocm-gfx1151, python-triton-gfx1151, python-aotriton-gfx1151	ROCm PyTorch 2.13, exact ROCm Triton 3.8, and AOTriton 0.13b in W2A
Inference engines	python-vllm-rocm-gfx1151, python-amd-aiter-gfx1151, python-flash-attn-rocm-gfx1151	TorchVision 0.28 and vLLM 0.27.1 in W2A; AITER 0.1.19.post2 stays experimental
Graph and quantization experiments	python-torchao-rocm-gfx1151, python-torch-migraphx-gfx1151	TorchAO 0.18.0 and Torch-MIGraphX ef3f901 remain nonblocking extensions
Model runners	llama.cpp-hip-gfx1151, llama.cpp-vulkan-gfx1151	llama.cpp b10369 HIP and Vulkan backends in the W2B Lemonade/llama.cpp work unit
Frontend and service layer	lemonade, lemonade-server, lemonade-app	Lemonade 11.5.2 waits on W0 source freeze before the W2B Lemonade/llama.cpp family build

Hand-maintained recipe packages carry package-local READMEs and recipe.json files with exact baselines, patch notes, and update hints. The generated TheRock package base has its own manifest and file-list evidence instead. Start at packages/ when you want package-specific details.

What Makes It Different

This repo is not a verbatim mirror of existing Arch or AUR packages. It carries Strix Halo-specific policy across the stack:

• gfx1151 ROCm targeting and TheRock split-package rendering
• amdclang and Zen 5 tuned build lanes where they are useful
• concrete llama.cpp HIP tuning such as -O3, -march=native, ThinLTO, -mprefer-vector-width=512, -famd-opt, gfx1151 HIP offload targeting, AMDGPU LLVM tuning, and optional LLVM inlining or Polly flags when the active compiler accepts them
• package-level replacements for runtime downloads, especially for Lemonade and llama.cpp
• Python 3.14 compatibility patches for packages whose upstream stable lanes are not there yet
• vLLM, AITER, FlashAttention, TorchAO, and Torch-MIGraphX patches that are tied to local ROCm behavior and host validation
• freshness and validation records that distinguish built, installed, passing, blocked, and merely advisory work

The patch overview is in docs/patches.md. Current validated state and known blockers are in docs/maintainers/current-state.md. Active candidate records point to their repo-local GitHub issues, with docs/backlog.md providing a compact discovery index.

The 2026-08 convergence route is issue-primary. Its blocking candidate line is CPython 3.14.7, TheRock 7.14, ROCm PyTorch 2.13 with exact ROCm Triton 3.8 and AOTriton 0.13b, the selected Transformers 5.15 dependency closure, TorchVision 0.28, vLLM 0.27.1, llama.cpp b10369 for both HIP and Vulkan, and the reviewed Lemonade fork selected after W0. These are candidates, not adoption claims; their exact gates and dispositions live in the linked execution issues and candidate ledger. Installed, built, deployed, smoked, and live-validated state remains evidence-owned by the maintainer docs.

Quick Start

Clone with submodules so the recipe input is available:

git clone --recurse-submodules https://github.com/nisavid/arch-strix-halo-pkgs.git
cd arch-strix-halo-pkgs

Then choose the path that matches your role.

I Want To Install Built Packages

This repository does not commit binary package archives. If repo/x86_64 has already been populated on your machine, publish it to a world-traversable local pacman repo and enable that repo in pacman:

sudo install -d /srv/pacman/strix-halo-gfx1151/x86_64
sudo rsync -a --delete repo/x86_64/ /srv/pacman/strix-halo-gfx1151/x86_64/

Then add the pacman stanza and install packages using the steps in the local repo guide. A typical install target is:

paru -S rocm-gfx1151 python-vllm-rocm-gfx1151 llama.cpp-hip-gfx1151 lemonade

I Want To Build Or Refresh Packages

Use ordinary package tools for a single package:

(cd packages/<pkgname> && makepkg -f)
python tools/update_pacman_repo.py \
  --package-dir packages/<pkgname> \
  --repo-dir repo/x86_64 \
  --require-packagelist

Use amerge when you want one command to plan, build, publish, and install a selected package set:

tools/amerge run python-vllm-rocm-gfx1151 llama.cpp-hip-gfx1151 lemonade

See Local Repo Usage for publish, repair, preview, resume, and inference-smoke workflows.

I Want To Understand The Repo

• Documentation index orients you by task and audience.
• Current state records evidence for what is installed, validated, or blocked.
• Backlog indexes the GitHub issues that own active work.
• Patch inventory explains the notable source changes.
• TheRock generator architecture explains the generated ROCm split packages.

Validation Story

The reference host has completed a full live cutover to the split ROCm stack and subsequent native rebuilds through tools/amerge. Tracked smokes cover ROCm visibility, Python imports, vLLM imports, llama.cpp entrypoints, Lemonade entrypoints, selected pooling/reranking flows, Qwen server routes, FlashAttention direct gates, and Torch-MIGraphX PT2E/Dynamo probes.

Some lanes are intentionally marked blocked or exploratory. For example, the FlashAttention CK direct package surface is validated, while the vLLM Qwen CK consumer path remains blocked in CK paged-KV behavior. The docs record that boundary so the next maintainer does not rediscover it from terminal scrollback.

Repository Map

Path	Purpose
packages/	Package directories, PKGBUILDs, package-local READMEs, recipes, and patches
policies/	Render and freshness policy used to keep package metadata reproducible
tools/	Repo maintenance tools, including amerge, package renderers, freshness checks, and inference smokes
generators/	TheRock split-package generation code
inference/scenarios/	Tracked local inference scenario catalog
docs/	User, maintainer, architecture, policy, evidence, and work-discovery documentation
upstream/ai-notes/	Blackcat Informatics recipe input submodule

Upstream And Baseline Inputs

This repo builds on work from:

• Blackcat Informatics' Strix Halo recipe work
• TheRock
• ROCm
• ROCm PyTorch
• ROCm Triton
• AOTriton
• AITER
• ROCm FlashAttention
• vLLM
• llama.cpp
• Lemonade
• Arch, CachyOS, and AUR package metadata used as package baselines